Editing images

Sometimes you don't want a brand new image — you want to change one thing on an image you already have: swap the background, remove an object, recolor a shirt, fix a typo on a banner. Edit image does exactly that, as a surgical follow-up to image generation rather than a fresh generation from scratch.

Where to find it

Look for the Edit image action (a wand icon) wherever a photo already lives in a creative:

• On any photo card in the creative's materials panel.
• On any photo variation inside a batch's Gallery view — next to the star and delete controls.
• On the primary reference image in the Picture step's Single picture tab.

Opening it starts the Edit image dialog with that photo as the source — nothing else in the creative is touched until you generate.

Choosing an edit type

Pick one of nine edit types, each tuned for a specific kind of change:

• Add object — add something new to a specific area.
• Remove object — remove something and naturally reconstruct what was behind it.
• Replace object — swap one object or area for another.
• Change background — change only the background, keep the subject and product unchanged.
• Clothing color — recolor a garment while preserving fit, texture, and everything else.
• Replace text — swap specific text while preserving font, layout, and surrounding design.
• Preserve identity — apply an edit while keeping a person or product's identity intact.
• Screenshot / mockup — edit an app screenshot or device mockup, keep the UI/frame intact.
• Offer card — edit an ad/offer card, keeping price, CTA, logo, and layout exact.

Picking a type pre-fills the Prompt field with a starting template — replace the [bracketed] placeholders with your specifics and edit the rest freely. It also resets the Preserve toggles below to sensible defaults for that edit type (for example, Offer card turns on Text, Logo, and Layout).

Mask vs. label targeting

Add, remove, replace, clothing color, and replace-text edits need to know where to apply the change — one of the two targeting methods below is required for them (the Target area field is marked accordingly until a mask is attached):

1. Draw a mask — choose Draw mask to paint directly over the area you want edited, using a brush and eraser with an adjustable size. Undo and Clear all are available while you draw. Confirm to attach the mask; a small preview and a remove (✕) action appear next to the button. A mask routes your edit to the most precise model available and unlocks all output formats and quality levels — use it whenever the edit needs to be exact, especially for text, logos, or product-safe areas.
2. Describe the target — type a short description into Target area instead (for example, "the empty table area" or "the woman's jacket"). This is best-effort: the model infers the region from your wording rather than an exact outline, so review the result carefully before accepting it.

You can also combine both — a mask with a label for extra context. Edit types that aren't location-specific (Change background, Preserve identity, Screenshot/mockup, Offer card) don't require either.

Preserve toggles

Below the target, a row of switches tells the model what to keep untouched: Identity/face, Product, Logo, Text, Layout, Colors, and — once a mask is attached — Outside mask. These seed sensible defaults per edit type, but adjust them for your specific image; turning more of them on generally means a more conservative, closer-to-original result.

Output controls and cost

Orientation, Quality, and Format adjust the output, though which ones are available depends on the route your edit predicts. Result name is pre-filled with the same conventional name other generations use (number, type, model, task, date) and tracks the predicted route until you edit it — replace it with anything you like, or clear it to let the backend assign a default.

Which output controls are available depends on the route:

• A drawn mask always routes to the most capable (and most expensive) model, with quality control and all three output formats (PNG, JPEG, WebP).
• A mask-free Change background edit routes to a lower-cost model that doesn't accept an explicit output format.
• Every other mask-free edit routes to a mid-cost, format-flexible model.

The dialog shows the estimated cost — in credits, or dollars if that's your display preference — and the predicted route right next to the Generate edit button before you submit, so you always know what a route costs before spending on it.

Reviewing the result

Once an edit finishes, it lands in the creative's materials panel just like a normal generation. Open the result and use Compare with source to drag a slider between the original and the edited image — or switch to the side-by-side toggle — to check that only the intended area changed.

From there you can:

• Re-run with same settings — resubmit with the same edit type, target, and route, only changing the prompt. Use this to nudge wording without redoing your mask.
• New edit — open a fresh Edit image dialog prefilled with the same edit type and prompt, targeting the result you just got. If the original edit used a mask, you'll need to draw it again — masks can't carry over between edits.

If a removal didn't come out clean, try expanding the mask slightly and generating again — a mask that's too tight around an object is the most common cause of visible leftovers.

Tips

• Prefer a drawn mask over a label whenever the edit touches text, a logo, or a product — label-only targeting is convenience, not precision.
• Turn on more Preserve toggles than you think you need; it's easier to relax them on a rerun than to fix an over-eager edit.
• Always use Compare with source before accepting a result — the point of Edit image is a surgical change, so anything different outside your target is worth a second look.
• For freeform transforms or combining several reference images instead of a targeted edit, use image generation's image-to-image flow instead.
• For a tour of every AI tool on a creative, see the AI generation overview.